Re: overhead of GC in caml runtime?

Re: overhead of GC in caml runtime?

[Damien Doligez]

>From: Norman Ramsey <nr@eecs.harvard.edu>

>Can anyone tell me approximately what fraction of time is
>spent in garbage collection, or even better, combined allocation and
>collection, in typical caml programs?  Or how to get caml to report
>this information for a particular program of mine?

As far as memory management is concerned, there is no typical caml
program, so you really need to get the information for your program.

GC timers have been on my to-do list for some time now but they are
going to involve non-negligible overhead, and I expect they'll have
rather poor accuracy.

In the meantime, you could compile your program with "ocamlopt -p",
guess from the names of the functions which ones belong to memory
management, and add the percentages by hand...

-- Damien

overhead of GC in caml runtime?

[Norman Ramsey]

Can anyone tell me approximately what fraction of time is
spent in garbage collection, or even better, combined allocation and
collection, in typical caml programs?  Or how to get caml to report
this information for a particular program of mine?

thanks,


Norman

Re: overhead of GC in caml runtime?

[Xavier Leroy]

> Can anyone tell me approximately what fraction of time is
> spent in garbage collection, or even better, combined allocation and
> collection, in typical caml programs?

It depends how allocation-intensive your program is.  For instance, the
Knuth-Bendix benchmark (which allocates quite a lot of short-lived
data) spends about 20% of its time in GC and allocation in the major
heap, when compiled with ocamlopt on a Pentium.

The percentage is lower for bytecode programs, because the collector
still runs at the same speed while the mutator runs more slowly
because of the interpretation overhead.

Most allocations in the minor heap are expanded in-line, so they can't
be measured.

Hope this helps,

- Xavier Leroy